SMFP

A molecule can be written as a SMILES string or drawn as a 2D graph, and the two are not redundant descriptions of the same thing. The string carries token order, branch syntax, and ring-closure bookkeeping that a chemist reads left to right; the graph carries adjacency, atom neighbourhoods, and substructure topology that no linear scan makes explicit. Pretrained molecular models have historically committed to one view or the other, and the multimodal ones that use both tend to encode each in its own tower and concatenate the two embeddings at the end. Concatenation is a weak form of fusion: each encoder learns in isolation, and a single downstream layer is left to work out which parts of which representation matter.

SMFP is a self-supervised pretraining framework that attacks that seam directly. Developed at Northwest Minzu University by Xingxing Li, Jing Jiang, and Yugang Dai, and published in Computational Biology and Chemistry in August 2026, it pairs a SMILES encoder with a molecular-graph encoder and inserts a bidirectional cross-attention module between them, so each modality queries the other while both are still being learned. A second device makes that fusion load-bearing rather than decorative: an adaptive masking strategy that places masks in non-overlapping positions across the tokenized SMILES and the graph. When a region is hidden in one view, the corresponding evidence is left visible in the other, and reconstructing it requires reading across the two representations instead of relying on within-modality context.

That non-overlapping masking idea is the distinguishing device of MoleSG; SMFP's own contribution is the bidirectional cross-attention fusion layered on top of it. The result sits alongside single-view pretrained models such as ChemBERTa as one pretrained representation transferred to many labelled tasks, rather than a method refit per dataset.

#Key Features

• Bidirectional cross-attention fusion: Sequence and graph representations attend to each other in both directions, letting the model weigh which modality carries the relevant signal for a given molecular region instead of concatenating fixed embeddings.
• Non-overlapping adaptive masking: Masked positions are allocated so that the SMILES and graph views hide different parts of the molecule, making cross-modal recovery the only route to the reconstruction objective.
• Self-supervised pretraining on unlabelled molecules: The representation is learned without property labels, which is what makes it transferable to the small labelled datasets typical of property prediction.
• One encoder, fourteen downstream tasks: The pretrained model is fine-tuned across eight classification and six regression benchmarks, covering both discrete endpoints and continuous targets.

#Technical Details

SMFP combines a transformer over tokenized SMILES with a graph encoder over the molecular graph, joined by the bidirectional cross-attention module that produces the fused representation, and is trained with a masked-reconstruction objective under the non-overlapping adaptive masking scheme. Evaluation uses fourteen datasets drawn from MoleculeNet — eight classification tasks and six regression tasks — with the pretrained encoder fine-tuned on each. Against representative published baselines, the authors report an advantage on 10 of the 14 datasets, which they characterise as competitive performance rather than a uniform win; the remaining four are cases where prior methods hold up. The paper is published under a subscription licence. Source code is public on GitHub — the SMILES encoder, the graph transformer, the cross-modal fusion block, the adaptive masking, and the pretraining and fine-tuning loops are all committed — but it carries no licence file and is not runnable as committed: there is no requirements file, no tokenizer or encoder configuration, and train_main.py imports Data_process.zinc_dataset_pretrain while the committed module is named zinc_pretrain_data.py. No pretrained checkpoint is released. The reported numbers therefore stand on the peer-reviewed evaluation rather than on independent reproduction.

#Applications

The target use is small-molecule property prediction in early drug discovery and cheminformatics: absorption and toxicity-style classification endpoints, solubility and other physicochemical regressions, and the general problem of scoring compounds when only a few hundred to a few thousand labelled examples exist. A pretrained multimodal encoder is most useful precisely in that regime, where a model trained from scratch on the labelled set alone overfits. Computational chemists screening virtual libraries, and groups building ADMET filters ahead of synthesis, are the direct beneficiaries.

#Impact

SMFP is a contribution to how molecular modalities are combined rather than to how much data they are pretrained on — an argument that the fusion mechanism, not the corpus size, is where multimodal molecular pretraining still has room to improve. Its claimed advantage on 10 of 14 MoleculeNet tasks is a modest, honestly framed result in a crowded benchmark suite where year-over-year gains are small and sensitive to splitting protocol. Uptake is impeded less by secrecy than by incompleteness: the article is closed access, the released code is unlicensed and does not run as committed, and no pretrained weights exist, so anyone building on SMFP will be repairing the repository and pretraining from scratch rather than loading a checkpoint. The cross-attention fusion module is the part of the design most likely to be carried into other multimodal molecular frameworks.